Files
OBLITERATUS/tests/test_persistence_pipeline.py
T

232 lines
7.8 KiB
Python

"""Integration contracts between checkpoint helpers and the model pipeline."""
from __future__ import annotations
import json
from pathlib import Path
from types import SimpleNamespace
from unittest.mock import MagicMock
import pytest
import torch
import obliteratus.abliterate as abliterate
from obliteratus.abliterate import AbliterationPipeline
pytestmark = pytest.mark.cpu
def _pipeline(tmp_path: Path) -> AbliterationPipeline:
pipeline = AbliterationPipeline(
model_name="test-model",
output_dir=str(tmp_path / "checkpoint"),
method="basic",
)
pipeline._on_log = lambda _message: None
pipeline._on_stage = lambda _event: None
pipeline.handle = MagicMock()
return pipeline
def test_rebirth_rejects_invalid_metadata_before_gathering_state(tmp_path):
pipeline = _pipeline(tmp_path)
destination = pipeline.output_dir
destination.mkdir()
sentinel = destination / "sentinel"
sentinel.write_text("old", encoding="utf-8")
pipeline._build_metadata = MagicMock(return_value={"invalid": object()})
pipeline._gather_state_dict = MagicMock()
with pytest.raises(TypeError):
pipeline._rebirth()
pipeline._gather_state_dict.assert_not_called()
assert sentinel.read_text(encoding="utf-8") == "old"
def test_rebirth_insufficient_capacity_preserves_checkpoint_and_runtime_state(
tmp_path,
monkeypatch,
):
pipeline = _pipeline(tmp_path)
destination = pipeline.output_dir
destination.mkdir()
sentinel = destination / "sentinel"
sentinel.write_text("old", encoding="utf-8")
pipeline._build_metadata = MagicMock(return_value={"schema": 1})
pipeline._gather_state_dict = MagicMock(
return_value={"weight": torch.ones(4, dtype=torch.float32)},
)
pipeline._write_local_checkpoint = MagicMock()
pipeline._free_gpu_memory = MagicMock()
pipeline._cleanup_offload_dir = MagicMock()
monkeypatch.setattr(
abliterate.shutil,
"disk_usage",
lambda _path: SimpleNamespace(free=16),
)
with pytest.raises(OSError, match="Insufficient disk space"):
pipeline._rebirth()
pipeline._write_local_checkpoint.assert_not_called()
pipeline._free_gpu_memory.assert_not_called()
pipeline._cleanup_offload_dir.assert_not_called()
assert sentinel.read_text(encoding="utf-8") == "old"
assert list(tmp_path.glob(".checkpoint.staging-*")) == []
def test_rebirth_ignores_non_os_disk_probe_failure_and_promotes_checkpoint(
tmp_path,
monkeypatch,
):
pipeline = _pipeline(tmp_path)
pipeline._build_metadata = MagicMock(return_value={"schema": 1})
state_dict = {"weight": torch.ones(1)}
pipeline._gather_state_dict = MagicMock(return_value=state_dict)
def write_checkpoint(checkpoint_dir, metadata_json, received_state_dict):
assert json.loads(metadata_json) == {"schema": 1}
assert received_state_dict is state_dict
(checkpoint_dir / "model.bin").write_bytes(b"saved")
pipeline._write_local_checkpoint = MagicMock(side_effect=write_checkpoint)
pipeline._free_gpu_memory = MagicMock()
pipeline._cleanup_offload_dir = MagicMock()
def fail_probe(_path):
raise RuntimeError("filesystem probe unavailable")
monkeypatch.setattr(abliterate.shutil, "disk_usage", fail_probe)
assert pipeline._rebirth() == pipeline.output_dir
assert (pipeline.output_dir / "model.bin").read_bytes() == b"saved"
pipeline._free_gpu_memory.assert_called_once_with()
pipeline._cleanup_offload_dir.assert_called_once_with()
def test_rebirth_pushes_only_after_local_checkpoint_promotion(tmp_path, monkeypatch):
import huggingface_hub
pipeline = _pipeline(tmp_path)
pipeline.push_to_hub = "auto"
pipeline.hub_token = "test-token"
pipeline.hub_community_org = "test-org"
pipeline._build_metadata = MagicMock(return_value={"schema": 1})
pipeline._gather_state_dict = MagicMock(return_value={"weight": torch.ones(1)})
pipeline._free_gpu_memory = MagicMock()
pipeline._cleanup_offload_dir = MagicMock()
def write_checkpoint(checkpoint_dir, _metadata_json, _state_dict):
(checkpoint_dir / "model.bin").write_bytes(b"saved")
pipeline._write_local_checkpoint = MagicMock(side_effect=write_checkpoint)
api = MagicMock()
api_factory = MagicMock(return_value=api)
monkeypatch.setattr(huggingface_hub, "HfApi", api_factory)
auto_name = MagicMock(return_value="test-org/test-model-OBLITERATED")
monkeypatch.setattr(abliterate, "auto_hub_repo_id", auto_name)
assert pipeline._rebirth() == pipeline.output_dir
assert (pipeline.output_dir / "model.bin").read_bytes() == b"saved"
api_factory.assert_called_once_with(token="test-token")
auto_name.assert_called_once_with(
"test-model",
api=api,
org="test-org",
)
api.create_repo.assert_called_once_with(
"test-org/test-model-OBLITERATED",
exist_ok=True,
)
api.upload_folder.assert_called_once_with(
folder_path=str(pipeline.output_dir),
repo_id="test-org/test-model-OBLITERATED",
commit_message="OBLITERATUS: abliterated test-model (basic)",
)
def test_rebirth_uses_fallback_token_for_explicit_hub_destination(
tmp_path,
monkeypatch,
):
import huggingface_hub
pipeline = _pipeline(tmp_path)
pipeline.push_to_hub = "test-org/explicit-repo"
pipeline._build_metadata = MagicMock(return_value={"schema": 1})
pipeline._gather_state_dict = MagicMock(return_value={"weight": torch.ones(1)})
pipeline._write_local_checkpoint = MagicMock(
side_effect=lambda path, *_args: (path / "model.bin").write_bytes(b"saved"),
)
pipeline._free_gpu_memory = MagicMock()
pipeline._cleanup_offload_dir = MagicMock()
api = MagicMock()
api_factory = MagicMock(return_value=api)
monkeypatch.setattr(huggingface_hub, "HfApi", api_factory)
monkeypatch.setenv("HF_TOKEN", "fallback-token")
pipeline._rebirth()
api_factory.assert_called_once_with(token="fallback-token")
api.create_repo.assert_called_once_with(
"test-org/explicit-repo",
exist_ok=True,
)
def test_write_local_checkpoint_strips_runtime_only_state_and_writes_metadata(
tmp_path,
monkeypatch,
):
import obliteratus.lora_ablation as lora_ablation
class Quantizer:
def __init__(self):
self.models = []
def remove_quantization_config(self, model):
self.models.append(model)
class Model:
def __init__(self):
self.hf_quantizer = Quantizer()
self._weight_conversions = {"legacy": "conversion"}
self.saved = None
def save_pretrained(self, path, **kwargs):
self.saved = (path, kwargs)
pipeline = _pipeline(tmp_path)
model = Model()
tokenizer = MagicMock()
pipeline.handle = SimpleNamespace(model=model, tokenizer=tokenizer)
pipeline._lora_adapters = {"layer": (torch.ones(1), torch.ones(1))}
checkpoint_dir = tmp_path / "staging"
adapter_path = checkpoint_dir / "lora"
save_adapters = MagicMock(return_value=adapter_path)
monkeypatch.setattr(lora_ablation, "save_lora_adapters", save_adapters)
state_dict = {"weight": torch.ones(1)}
metadata_json = '{"schema": 1}'
checkpoint_dir.mkdir()
pipeline._write_local_checkpoint(checkpoint_dir, metadata_json, state_dict)
assert model.hf_quantizer.models == [model]
assert not hasattr(model, "_weight_conversions")
assert model.saved == (
checkpoint_dir,
{
"state_dict": state_dict,
"max_shard_size": "2GB",
"save_original_format": False,
},
)
tokenizer.save_pretrained.assert_called_once_with(checkpoint_dir)
save_adapters.assert_called_once_with(pipeline._lora_adapters, checkpoint_dir)
assert (checkpoint_dir / "abliteration_metadata.json").read_text(
encoding="utf-8",
) == metadata_json